Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Date:   Tue, 22 Sep 2020 13:29:17 -0700
From:   "Paul E. McKenney" <paulmck@kernel.org>
To:     Neeraj Upadhyay <neeraju@codeaurora.org>
Cc:     josh@joshtriplett.org, rostedt@goodmis.org,
        mathieu.desnoyers@efficios.com, jiangshanlai@gmail.com,
        joel@joelfernandes.org, rcu@vger.kernel.org,
        linux-kernel@vger.kernel.org
Subject: Re: [PATCH] rcu/tree: Correctly handle single cpu check in
 rcu_blocking_is_gp
Message-ID: <20200922202917.GB29330@paulmck-ThinkPad-P72>
Reply-To: paulmck@kernel.org
References: <1600717557-19880-1-git-send-email-neeraju@codeaurora.org>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <1600717557-19880-1-git-send-email-neeraju@codeaurora.org>
User-Agent: Mutt/1.9.4 (2018-02-28)
Precedence: bulk

On Tue, Sep 22, 2020 at 01:15:57AM +0530, Neeraj Upadhyay wrote:
> Currently, for non-preempt kernels (with CONFIG_PREEMPTION=n),
> rcu_blocking_is_gp() checks (with preemption disabled), whether
> there is only one cpu online. It uses num_online_cpus() to
> decide whether only one cpu is online. If there is only single
> cpu online, synchronize_rcu() is optimized to return without
> doing all the work to wait for grace period. However, there are
> few issues with the num_online_cpus() check used:

Great catch!!!

I do have some questions about your suggested fix, though.

> 1. num_online_cpus() does a atomic_read(&__num_online_cpus). As
>    hotplug locks are not held, this does not ensure that
>    new incoming cpus update of the count is visible. This can
>    result in read side section on new incoming cpu, observe
>    updates which should not be visible beyond the grace period
>    corresponding to synchronize_rcu().
> 
>    For e.g. below litmus test, where P0 process corresponds to
>    synchronize_rcu() and P1 corresponds to new online cpu,  has
>    positive witnesses; confirming the possibility of read side
>    section to extend before and after the grace period, thereby
>    breaking guarantees provided by synchronize_rcu().
> 
>    {
>      int x = 0;
>      atomic_t numonline = ATOMIC_INIT(1);
>    }
> 
>    P0(int *x, atomic_t *numonline)
>    {
>      int r0;
>      WRITE_ONCE(*x, 1);
>      r0 = atomic_read(numonline);
>      if (r0 == 1) {
>        smp_mb();
>      } else {
>        synchronize_rcu();
>      }
>      WRITE_ONCE(*x, 2);
>    }
> 
>    P1(int *x, atomic_t *numonline)
>    {
>      int r0; int r1;
> 
>      atomic_inc(numonline);
>      smp_mb();
>      rcu_read_lock();
>      r0 = READ_ONCE(*x);
>      smp_rmb();
>      r1 = READ_ONCE(*x);
>      rcu_read_unlock();
>    }
> 
>    locations [x;numonline;]
> 
>    exists (1:r0=0 /\ 1:r1=2)
>
> 2. Second problem is, the same early exit, from synchronize_rcu()
>    does not provide full ordering, memory barrier, w.r.t. memory
>    accesses after synchronize_rcu() call.
> 
> 3. Third, more important issue is related to outgoing cpu. Checking
>    only for __num_online_cpus with preemotion disabled isn't sufficient
>    for RCU, as, on completion of CPUHP_TEARDOWN_CPU stop machine (which
>    clears outgoing cpu from __num_online_cpus, the CPU switches to idle
>    task. So, checking only for __num_online_cpus does not consider
>    RCU read side sections in scheduler code (before switching to idle
>    task) and any potential read side sections in idle task, before final
>    RCU-quiesce entry into cpuhp_report_idle_dead() -> rcu_report_dead().
> 
> To handle these issues, add a new rcu_state member n_online_cpus, to
> keep account of the current number of online cpus. The atomic updates
> to this counter from rcu_report_dead() and rcu_cpu_starting() and
> the added read/write memory ordering semantics ensure that
> synchronize_rcu() fast path waits for all read side sections, where
> incoming/outgoing cpus are considered online, for RCU i.e. after
> rcu_cpu_starting() and before rcu_report_dead().
> 
> Signed-off-by: Neeraj Upadhyay <neeraju@codeaurora.org>
> ---
> 
> Below is the reproducer for issue described in point 3; this snippet
> is based on klitmus generated test, which is modified to sample reads
> from idle thread:
> 
> static void code0(int* x) {
> WRITE_ONCE(*x, 1);
>   idle_ctr = 0;
>   smp_mb();
>   mdelay(10);
>   WRITE_ONCE(*x, 1);
>   idle_ctr = 1;
>   synchronize_rcu();
>   WRITE_ONCE(*x, 2);
>   idle_ctr = 2;
> 
> }
> 
> static int thread0(void *_p) {
>   int _j, _i;
>   ctx_t *_a = (ctx_t *)_p;
> 
>   smp_mb();
>   for (_j = 0 ; _j < stride ; _j++) {
>     for (_i = _j ; _i < size ; _i += stride) {
>       while (idle_wait1) {
>           cpu_relax();
>           cond_resched();
>         }
>       code0(&_a->x[_i]);
>       smp_mb();
>       get_online_cpus();
>       idle_wait1 = true;
>       put_online_cpus();
>     }
>   }
>   atomic_inc(&done);
>   smp_mb();
>   wake_up(wq);
>   smp_mb();
>   do_exit(0);
> }
> 
> 
> static void code1(int* x,int* out_1_r1,int* out_1_r0) {
> 
>   int r0; int r1;
> 
>   r0 = READ_ONCE(idle_ctr_snap1);
>   r1 = READ_ONCE(idle_ctr_snap2);
> 
>   *out_1_r1 = (int)r1;
>   *out_1_r0 = (int)r0;
> }
> 
> static int thread1(void *_p) {
>   ctx_t *_a = (ctx_t *)_p;
>   int _j, _i;
> 
>   smp_mb();
>   for (_j = 0 ; _j < stride ; _j++) {
>     for (_i = _j ; _i < size ; _i += stride) {
>       while (idle_wait2) {
>           cpu_relax();
>           cond_resched();
>       }
>       get_online_cpus();
>       code1(&_a->x[_i],&_a->out_1_r1[_i],&_a->out_1_r0[_i]);
>       smp_mb();
>       idle_wait2 = true;
>       put_online_cpus();
>     }
>   }
>   atomic_inc(&done);
>   smp_mb();
>   wake_up(wq);
>   smp_mb();
>   do_exit(0);
> }
> 
> Idle thread snippet:
> 
> if (cpu_is_offline(cpu)) {
>         smp_mb();
>         idle_wait1 = false;
>         mdelay(8);
>         smp_mb();
>         rcu_read_lock();
>         idle_ctr_snap1 = idle_ctr;
>         mdelay(40);
>         smp_rmb();
>         idle_ctr_snap2 = idle_ctr;
>         rcu_read_unlock();
>         smp_mb();
>         idle_wait2 = false;
>         tick_nohz_idle_stop_tick();
>         cpuhp_report_idle_dead();
>         arch_cpu_idle_dead();
> 
>  kernel/rcu/tree.c | 65 +++++++++++++++++++++++++++++++++++++++++++++++++++++++
>  kernel/rcu/tree.h |  1 +
>  2 files changed, 66 insertions(+)
> 
> diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
> index 2424e2a..33493f0 100644
> --- a/kernel/rcu/tree.c
> +++ b/kernel/rcu/tree.c
> @@ -3609,9 +3609,59 @@ static int rcu_blocking_is_gp(void)
>  	if (IS_ENABLED(CONFIG_PREEMPTION))
>  		return rcu_scheduler_active == RCU_SCHEDULER_INACTIVE;
>  	might_sleep();  /* Check for RCU read-side critical section. */
> +	/*
> +	 * a = p
> +	 * a = NULL
> +	 * synchronize_rcu()
> +	 *   rcu_blocking_is_gp()
> +	 *     num_online_cpus()
> +	 *       atomic_read(&__num_online_cpus)
> +	 * kfree(p);
> +	 *
> +	 *      - VS -
> +	 *
> +	 *              __cpu_up()
> +	 *                set_cpu_online(cpu)
> +	 *                  atomic_inc(&__num_online_cpus)
> +	 *              rcu_read_lock()
> +	 *              rcu_dereference(a) (a == p)
> +	 *              rcu_read_unlock()
> +	 *
> +	 * rcu_blocking_is_gp() must observe atomic_inc(&__num_online_cpus),
> +	 * in order to ensure that, RCU read side critical section on new
> +	 * online cpu, either start after synchronize_rcu()'s GP starts or
> +	 * it completes before synchronize_rcu() returns.
> +	 *
> +	 * However, atomic_read(&__num_online_cpus) does not ensure that.
> +	 *
> +	 * Essentially, below condition exist:
> +	 *
> +	 * {
> +	 *   int x = 0;
> +	 *   atomic_t numonline = ATOMIC_INIT(1);
> +	 * }
> +	 *
> +	 * P0(int *x, atomic_t *numonline)     P1(int *x, atomic_t *numonline)
> +	 * {                                   {
> +	 *   int r0;                             int r0; int r1;
> +	 *   WRITE_ONCE(*x, 1);                  atomic_inc(numonline);
> +	 *   r0 = atomic_read(numonline);        rcu_read_lock();
> +	 *   if (r0 == 2) {                      r0 = READ_ONCE(*x);
> +	 *     synchronize_rcu();                smp_rmb();
> +	 *   }                                   r1 = READ_ONCE(*x);
> +	 *   WRITE_ONCE(*x, 2);                  rcu_read_unlock();
> +	 * }                                   }
> +	 *
> +	 * exists (1:r0=0 /\ 1:r1=2)
> +	 *
> +	 * atomic_add_return(0, &rcu_state.n_online_cpus) and corresponding
> +	 * atomic_inc(&rcu_state.n_online_cpus) in rcu_cpu_starting() corrects
> +	 * this ordering issue.
> +	 */
>  	preempt_disable();
>  	ret = num_online_cpus() <= 1;

Here I assume that rcu_state.n_online_cpus is incremented early in
the CPU-hotplug CPU-online process, that is, on one of the CPUs that
was running prior to the new CPU coming online.  (The problem with the
existing code is not the lack of ordering, but rather that the changes
to the number of online CPUs happen in places that are not helpful to
synchronize_rcu().)

If rcu_state.n_online_cpus is equal to one at any point in this region of
code, there is only one CPU, and that CPU sees all prior accesses made
by any CPU that was online at the time of its access.  Furthermore, if
rcu_state.n_online_cpus is equal to one, its value cannot change until
after the preempt_enable() below.

Furthermore, if n_online_cpus is equal to one here, all later CPUs
(both this one and any that come online later on) are guaranteed to see
all accesses by any CPU prior to this point in the code, and without
added memory barriers.  Those memory barriers have to be present in the
CPU-hotplug code or lots of things would break.

On the other hand, if n_online_cpus is greater than one, then we
will be using the heavyweight call to synchronize_rcu(), which will
guarantee all the ordering we need.  (Please refer to the rather
lengthy header comment for synchronize_rcu().)

So if you access rcu_state.n_online_cpus with preemption disabled,
READ_ONCE() suffices and no memory barriers are required.

>  	preempt_enable();

And we only get to this point in the code when CONFIG_PREEMPT_NONE=y,
so the preempt_disable() and preempt_enable() are optional.  Though they
can be argued to be useful documentation.  Or maybe not...

> +	ret = ret && (atomic_add_return(0, &rcu_state.n_online_cpus) <= 1);
>  	return ret;
>  }
>  
> @@ -3655,6 +3705,11 @@ void synchronize_rcu(void)
>  			 lock_is_held(&rcu_sched_lock_map),
>  			 "Illegal synchronize_rcu() in RCU read-side critical section");
>  	if (rcu_blocking_is_gp())
> +		/*
> +		 * atomic_add_return() in rcu_blocking_is_gp () provides
> +		 * full memory barrier ordering with any rcu section after
> +		 * synchronize_rcu() call.
> +		 */

Given your fix of having RCU keep its own count of the number of online
CPUs, no additional ordering is required.  Either synchronize_rcu()
provides what is required or we are in single-CPU state, meaning we
don't need any ordering.

>  		return;
>  	if (rcu_gp_is_expedited())
>  		synchronize_rcu_expedited();
> @@ -4086,6 +4141,10 @@ void rcu_cpu_starting(unsigned int cpu)
>  	mask = rdp->grpmask;
>  	raw_spin_lock_irqsave_rcu_node(rnp, flags);
>  	WRITE_ONCE(rnp->qsmaskinitnext, rnp->qsmaskinitnext | mask);
> +	/* Order with access of n_online_cpus in rcu_blocking_is_gp */
> +	atomic_inc(&rcu_state.n_online_cpus);
> +	/* Order with rcu-side usages after this */
> +	smp_mb__after_atomic();

Ah, here is the problem.  Please instead put the increment in
rcutree_prepare_cpu(), which in the one-to-two transition will be running
on the single CPU in the system, thus avoiding the need for ordering.
Yes, this will result in unnecessary calls to synchronize_rcu() during
the CPU-online process, but who cares?  ;-)

>  	newcpu = !(rnp->expmaskinitnext & mask);
>  	rnp->expmaskinitnext |= mask;
>  	/* Allow lockless access for expedited grace periods. */
> @@ -4138,6 +4197,12 @@ void rcu_report_dead(unsigned int cpu)
>  		raw_spin_lock_irqsave_rcu_node(rnp, flags);
>  	}
>  	WRITE_ONCE(rnp->qsmaskinitnext, rnp->qsmaskinitnext & ~mask);
> +	/*
> +	 * Order with access of n_online_cpus in rcu_blocking_is_gp().
> +	 * Release semantics ensures that RCU read sections before it
> +	 * are observed by rcu_blocking_is_gp().
> +	 */
> +	atomic_dec_return_release(&rcu_state.n_online_cpus);

Similarly, please put this decrement into rcutree_dead_cpu(), which
runs on one of the remaining CPUs after the outgoing CPU is long gone.
In the two-to-one transition, this will run on the single remaining
CPU in the system, thus avoiding the need for ordering.  Again, yes,
this will result in unnecessary calls to synchronize_rcu() during the
CPU-online process, but again who cares?

>  	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
>  	raw_spin_unlock(&rcu_state.ofl_lock);
>  
> diff --git a/kernel/rcu/tree.h b/kernel/rcu/tree.h
> index e4f66b8..4d9a9c0 100644
> --- a/kernel/rcu/tree.h
> +++ b/kernel/rcu/tree.h
> @@ -298,6 +298,7 @@ struct rcu_state {
>  						/* Hierarchy levels (+1 to */
>  						/*  shut bogus gcc warning) */
>  	int ncpus;				/* # CPUs seen so far. */
> +	atomic_t n_online_cpus;			/* # CPUs online for RCU. */

With those changes in place, this can be just an int.  The increments
and decrements can use normal C-language loads and WRITE_ONCE() for the
stores.  The trick is that this value will only change from one to two
(and vice versa) when there is only one online CPU.

And the num_online_cpus() can be replaced with a READ_ONCE().

Does this make sense, or am I missing something?

							Thanx, Paul

>  	/* The following fields are guarded by the root rcu_node's lock. */
>  
> -- 
> The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,
> a Linux Foundation Collaborative Project
>